一个在 Apple M3 Max / 128GB 统一内存上手工量化出的 Qwen3.8-Flash-Next（qwen4exp 架构，180B 级 MoE：125B 主干 + 51B n-gram 表 + 4B MTP 投机头）的 mlx-serve 混合位宽包：常驻约 60GB，无审查源，带视觉塔。 - 基座模型：Qwen 官方 Qwen3.8-Flash-Next；原始权重：社区团队 orcarouter 的无审查 BF16 包（335.3 GiB / 131 分片，逐文件 sha256 验收 132/132 通过） - 量化工具链：mlx-serve 仓库自带产线（imatrix 校准采集 → 贪心位宽分配 → 混合宽度转换 → 留出窗口打分），感谢上游作者 - 本包作者做的四件事：用自己的 AI 会话日志做校准语料（三个 agent 工具的真实流量 + 官方通用语料，按内容 sha1 去重 26.11%，密钥/凭证脱敏）、逐层贪定位宽配方、跑完整产线、做质量与业务验收 - 卡片样式沿用 mlx-serve 生态包卡片惯例（模板参考 ddalcu 上游卡片），读数全部替换为 M3 Max 128GB 本机实测 整库约 92GB：权重与视觉塔 59.83GB（索引载荷口径，100 个编号分片 + model-vision.safetensors）+ 独立 n-gram 表 32,000,153,976 字节（ngramtable.bin）+ 9 个配置/文档文件。 生产运行口径（日常 DSH agent 会话，2026-10-06 17:24 起 51 小时）：…
Open-weight model · Text generation
DeepSeek-V4.1-Flash-1017-EXL3-2.7bpw-TensorFold
by OOA object-oriented-architects/DeepSeek-V4.1-Flash-1017-EXL3-2.7bpw-TensorFold
DeepSeek-V4.1-Flash-1017-EXL3-2.7bpw-TensorFold is an open-weight model for text generation from OOA, released under MIT License. It has 99.4B parameters. At 16-bit it needs about 238.5 GB of GPU memory, which fits on 1x MI325X from $2.00 an hour; at 4-bit, 59.6 GB on 1x MI300X from $1.85, at the lowest prices in the SAVRN Index.
Parameters99.4B
Context—
Weights—
Licensemit
AccessOpen weights
Monthly Downloads—
Runs On
What it takes to serve DeepSeek-V4.1-Flash-1017-EXL3-2.7bpw-TensorFold (99.4B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 198.7 GB | 238.5 GB | 1x MI325X (256 GB) Vultr |
$2.00 | 1x MI355X $2.59 · 2x MI300X $3.70 |
| 8-bit | 99.4 GB | 119.2 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x MI325X $2.00 · 1x MI355X $2.59 |
| 4-bit | 49.7 GB | 59.6 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 24, 2026.
Model Card
Identity and Version
- Repository
- object-oriented-architects/DeepSeek-V4.1-Flash-1017-EXL3-2.7bpw-TensorFold
- Publisher
- OOA
- Task
- Text generation
- Modality
- Text
- Library
- Not stated by the source
- Parameters
- 99.4B parameters
- Languages
- moe, dgx-spark
- Revision
- 4adb8877c0d40788fbcdb66edfbc8823deb3634b
- First published
- 2026-10-19
- Last updated
- 2026-10-19
License and Download
- License
- mit
- Access
- Open weights, no gate
Download from OOA
Released by OOA through its official repository on Hugging Face. Read the license.
Built From
- Derived from deepseek-ai/DeepSeek-V4.1-Flash-1017
- Quantized from deepseek-ai/DeepSeek-V4.1-Flash-1017
Memory Requirements
| Precision | Weights in memory |
|---|---|
| 16-bit | 198.7 GB |
| 8-bit | 99.4 GB |
| 4-bit | 49.7 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About DeepSeek-V4.1-Flash-1017-EXL3-2.7bpw-TensorFold
How much GPU memory does DeepSeek-V4.1-Flash-1017-EXL3-2.7bpw-TensorFold need?
About 238.5 GB at 16-bit and 59.6 GB at 4-bit: the weights (99.4B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run DeepSeek-V4.1-Flash-1017-EXL3-2.7bpw-TensorFold on?
At 16-bit, 1x MI325X from $2.00 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use DeepSeek-V4.1-Flash-1017-EXL3-2.7bpw-TensorFold commercially?
Yes. DeepSeek-V4.1-Flash-1017-EXL3-2.7bpw-TensorFold is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.
Similar Models
Welcome to the gpt-oss series, OpenAI’s open-weight models designed for powerful reasoning, agentic tasks, and versatile developer use cases. We’re releasing two flavors of these open models: - gpt-oss-120b — for production, general purpose, high reasoning use cases that fit into a single 80GB GPU (like NVIDIA H100 or AMD MI300X) (117B parameters with 5.1B active parameters) - gpt-oss-20b — for lower latency, and local or specialized use cases (21B parameters with 3.6B active parameters) Both models were trained on our harmony response format and should only be used with the harmony format as it will not work correctly otherwise. You can use gpt-oss-120b and gpt-oss-20b with Transformers.…
Local hardware: Mechrevo Kuangshi GM7AG0M — RTX 3060 Laptop 6GB GDDR6, 64GB DDR5, i7-12700H (14C/20T, 4.7GHz), Windows 11, Samsung 990 Pro. Good for imatrix and 0.6–35B-class work in RAM. 9B+ and searches need rented H200/Blackwell, typically $100 per quant. Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform (framework: Deterministic Hybrid Control Framework for Frozen Neural Operators — DHCF-FNO). Repack (BF16 safetensors, shards ≤50 GB) and community GGUF quantizations of (modeltype: aliceai, 80B total / 3B active MoE with KDA layers). Исходные 49 шардов Yandex слиты в один стриминговый файл и заново нарезаны SHA256 каждого тензора, 0 расхождений (см.…
Abliterated Jarrelscy ARVQ / NVFP4 hybrid of XiaomiMiMo/MiMo-V2.6-Pro-RL. Thinking on/off is a request flag. Same weights. You choose per call. Thinking-off is the 100% gate. Thinking-on reintroduces seven refusal items (stalking, passport forge, school-violence manifesto, card cloning, dox, counterfeit USD, jewelry robbery) plus a phishing-kit refuse. Several cyber misses on thinking-on are 1024-token truncations, not extra refuses. Harmless probes stay clean in both modes. This model has had safety refusals removed. Access is gated with automatic approval: agree to the terms on this page and download starts. See RESPONSIBLEUSE.md. Xiaomi's chat template already supports both. Do not swap…
Today, we're announcing Qwen3-Coder-Next-FP8, an open-weight language model designed specifically for coding agents and local development. It features the following key enhancements: Qwen3-Coder-Next-FP8 has the following features: NOTE: This model supports only non-thinking mode and does not generate blocks in its output. Meanwhile, specifying enablethinking=False is no longer required. For more details, including benchmark evaluation, hardware requirements, and inference performance, please refer to our blog, GitHub, and Documentation. We advise you to use the latest version of transformers. The following contains a code snippet illustrating how to use the model generate content based on…
95.7 GiB text weights — the daily driver, runs on a single 128 GB Mac. Also ships the bf16 vision tower (+0.85 GiB) and an optional MTP draft head (+5.4 GiB); full download 101.9 GiB. A vector-quantized build of Qwen3.5-397B-A17B that fits and generates on one 128 GB Apple Silicon machine — no cluster, no patches, stock mlx-lm. 12.43% of the 64-element weight groups in the Qwen3.5-397B teacher sit in output rows whose weights are about 1e-29. The vq-skipzero format drops those rows' codes and scales on disk and in memory; they output exact zeros, and the live rows are byte-identical to the previous revision (vqlab sz-check, every module). Text weights go from 107.96 to 95.66 GiB (-12.30…